Computer-implemented method for determining a control function

Abstract

The invention relates to a computer-implemented method (100) for determining a control function (28) for predicting at least one state value (22) of a vehicle component, wherein the method (100) comprises at least the following steps: determining (102) at least one internal parameter (24) of a learning algorithm (20) that has been trained by changing the at least one internal parameter (24) to provide at least one output signal that indicates at least one state value (22) of the vehicle component from at least one input signal that indicates at least one parameter of the vehicle component (14-18); and determining (104) a control function (28) for predicting the state value (22) of the vehicle component by means of a symbolic regression based at least on the input signal and the internal parameter (24) of the trained learning algorithm (20).With the method (100) the costs and computational effort for determining the state values (22) can be reduced.

Description

[0001] The invention relates to a computer-implemented method for determining a control function.

[0002] The simulation or prediction of condition values of vehicle components can be performed for vehicle operation, for example, in autonomous driving, or during vehicle design to analyze their behavior in advance. Artificial neural networks can be used for such simulations, as is known, for example, from DE 10 2020 212 280 A1. These neural networks are trained to predict the corresponding condition values from specific measured values of the vehicle component. However, neural networks require very large computing resources to run.

[0003] From DE 10 2022 108 459 B3, it is known to create and adapt a formula for calculating emission values of an internal combustion engine based on measured emission values using symbolic regression according to Udrescu, S.; Tegmark, M.: Al Feynman: A physics-inspired method for symbolic regression. In: Science Advances, Vol. 6, 2020, No. 16, pp. 1-16. - ISSN 2375-25, DOI: 10.1126 / sciadv.aay2631 based on a neural network.

[0004] The object of the invention is to provide a method by which the costs and computational effort for determining the state values are reduced.

[0005] The problem is solved by the features of the independent claims. Advantageous further developments are the subject of the dependent claims and the following description.

[0006] According to a first aspect, a computer-implemented method for determining a control function for predicting at least one state value of a vehicle component is described, wherein the method comprises at least the following steps: determining at least one internal parameter of a learning algorithm that has at least one neural network with a fluid time constant, such that it is designed as an LTC network, and that has been trained by changing the at least one internal parameter to provide at least one output signal that indicates at least one state value of the vehicle component from at least one input signal that indicates at least one parameter of the vehicle component; and determining a control function for predicting the state value of the vehicle component by means of symbolic regression based at least on the input signal and the internal parameter of the trained learning algorithm.

[0007] This provides a computer-implemented method that provides a control function for predicting state values of vehicle components that requires comparatively few computing resources and thus saves costs. For this purpose, a self-learning algorithm, which may be a neural network, for example, is first trained to determine a state value of the vehicle component from vehicle parameters. During training, the self-learning algorithm is adapted to the desired goal by changing its internal parameters. At least the internal parameters and the input signal of the trained self-learning algorithm are then used to determine a mathematical function as a control function for predicting state values of the vehicle component using symbolic regression.Since the control function merely links input values using mathematical operations and an output value is immediately determined from this, a state value for the vehicle component can be determined from parameters of the vehicle component or the vehicle with comparatively few mathematical operations.

[0008] According to some embodiments, it is conceivable that the learning algorithm may comprise at least one neural network with at least one neural circuit.

[0009] With appropriate training, neural networks can recognize patterns in input values and provide corresponding output values. This allows neural networks, after training, to make predictions about, in this case, state values with high accuracy. Accordingly, the control function provided by symbolic regression will also be highly accurate.

[0010] The neural network is designed as a liquid time constant neural network (LTC network). In an LTC network, the data flow in the hidden layers is calculated, among other things, using a system of linear differential equations. An LTC network with just a few hundred nodes can recognize patterns in time-varying parameters with high accuracy. The number of nodes in the LTC network can therefore be significantly lower than the number of nodes in conventional neural networks, which can typically have thousands to millions of nodes. This allows an LTC network to have comparatively few internal parameters, simplifying symbolic registration. Furthermore, LTC networks can be more robust against fluctuations in input values than conventional neural networks.

[0011] According to some embodiments, it is conceivable that the parameter of the vehicle component can have at least one feature of at least one labeled real measured value from the vehicle.

[0012] This allows real measurements from the vehicle to be used to train the adaptive algorithm. This increases the likelihood that the adaptive algorithm will receive data for training that covers the entire range of vehicle component parameters. Accordingly, symbolic regression will be able to determine a control function that delivers the same or nearly the same results as the adaptive algorithm for the entire range provided.

[0013] According to some embodiments, it is conceivable that the at least one feature may be an element of the group: input of a control logic, output of a control logic, mathematical operator, name of a variable and / or physical unit.

[0014] With these features, the accuracy of the control function can be further increased when determined by symbolic regression.

[0015] According to some embodiments, it is conceivable that the method, before determining the at least one internal parameter of the trained adaptive algorithm, can further comprise at least the following step: training an adaptive algorithm with a training data set which has at least measurement data of at least one parameter of the vehicle component as a training input signal and measurement data of the at least one state value of the vehicle component, which are assigned to the measurement data of the parameters, as a target output signal, wherein during training the at least one internal parameter of the adaptive algorithm is changed in order to adapt an output signal of the adaptive algorithm based on the training input signal to the target output signal.

[0016] The training dataset can be compiled manually by a user in advance. Furthermore, the training dataset can be prepared in such a way that it covers the entire, or almost the entire, range that the corresponding parameters of the vehicle component or vehicle can assume.

[0017] According to some embodiments, it is conceivable that after training the learning algorithm, the internal parameters can be fixed.

[0018] This prevents the internal parameters from changing further as the adaptive algorithm receives additional input signals. This establishes a point in time at which it is determined that the accuracy of the adaptive algorithm is sufficient. To prevent any possible deterioration in accuracy, the internal parameters can then be fixed.

[0019] According to some embodiments, it is conceivable that the vehicle component can be a battery of a battery electric vehicle and the state value can be a battery voltage.

[0020] Thus, the method explained above can be used to provide a control function for a battery of a battery-electric vehicle that predicts the battery voltage depending on the parameters of the battery or the vehicle.

[0021] According to some embodiments, it is conceivable that the step of determining a control function can be performed using a software tool for performing symbolic regression.

[0022] For example, a tool based on the Python programming language can be used for symbolic regression. This allows a modern tool to be used to perform symbolic registration, which can provide a high-quality control function.

[0023] According to some embodiments, it is conceivable that the determined control function can be used during operation of the vehicle and / or during construction of the vehicle to predict the at least one state value of the vehicle component.

[0024] The control function allows status values for vehicle components to be determined quickly or in real time with high accuracy. This can improve the quality of control by an autonomous driving system or the quality of the vehicle's design.

[0025] According to a further aspect, a computer program product is described, comprising instructions which, when the program is executed by a computer, cause the computer to carry out the steps of the method according to the preceding description.

[0026] Advantages and effects, as well as further developments of the computer program product, arise from the advantages and effects, as well as further developments of the method described above. Reference is therefore made to the preceding description in this regard. A computer program product can be understood, for example, as a data storage medium on which a computer program element is stored that contains instructions executable by a computer. Alternatively or additionally, a computer program product can also be understood, for example, as a permanent or volatile data storage device, such as flash memory or RAM, that contains the computer program element. However, this does not exclude other types of data storage devices that contain the computer program element.

[0027] The invention is described below using an exemplary embodiment with reference to the accompanying drawings. They show: Fig. 1 a flow chart of the process; Fig. 2 a schematic representation of measurement data for the vehicle component; Fig. 3 a schematic representation of the training of the learning algorithm; Fig. 4 a schematic representation of symbolic regression; and Fig. 5 a schematic representation of the use of the control function.

[0028] According to Fig. 1, the computer-implemented method in its entirety is referred to below by the reference numeral 100.

[0029] The computer-implemented method 100 uses a learning algorithm that can be first trained according to the optional step 106.

[0030] The adaptive algorithm can be trained using a training data set that can include measurement data of at least one parameter of the vehicle component as the training input signal. The training data set can include measurement data of at least one state value of the vehicle component as the target output signal.

[0031] The measurement data are exemplary in Fig. 2. From a database 10 containing measurement data of a vehicle or complete vehicle components, at least one data set can be used, which is represented in visualized form with the reference symbol 12. The visualized form of the data set 12 can display names 14 of the measurement data, temporal profiles 16 of the measurement data, and / or values 18 of the measurement data. The temporal profiles 16 of the measurement data can also include, among other things, the measured condition values of the vehicle component.

[0032] According to Fig. 3, the adaptive algorithm can be designated by reference numeral 20 and can be implemented as a neural network. The adaptive algorithm 20 is expediently implemented as a neural network with a fluid time constant. This means that the data flows between the hidden layers of the neural network are described, among other things, by ordinary differential equations. This means that the temporal change in the data flows can depend on the time-dependent variables of the input signal. This can introduce nonlinearity into the data flows, which can make the results of the neural network more robust against changes in the input values.

[0033] The training of the adaptive algorithm 20 is achieved by changing internal parameters 24 of the adaptive algorithm 20 until the adaptive algorithm 20 provides an output signal that substantially corresponds to the desired output signal. The goal of the training is for the adaptive algorithm 20 to be able to predict a state value 22 for a vehicle component from the input signal, which may indicate parameters 14-18.

[0034] Once training is complete, the internal parameters 24 of the adaptive algorithm can be fixed in a further optional step 108. This means that the internal parameters 24 should no longer be changed.

[0035] In a further step 102, the internal parameters 24 of the adaptive algorithm 20 are determined. This can be done by a distillation process or extraction process, especially if the adaptive algorithm 20 comprises a neural network.

[0036] Using the internal parameters 24 and the input signal, which was used, for example, for training the adaptive algorithm 20, a control function 28 can be determined by means of symbolic regression in a further step 104. The control function 28 can be configured to predict the state value 22 of the vehicle component.

[0037] The determination of the control function 28 is exemplified in Fig. 4. A software tool 26 can be used to perform the symbolic regression. The software tool 26 can be based on the Python programming language, for example. However, any other software tool 26 capable of performing symbolic regression is also conceivable.

[0038] In a further optional step 110, the determined control function 28 can be used during the operation of a vehicle or during the construction of a vehicle for the determination of state values 22 of a vehicle component.

[0039] This is exemplified in Fig. 5. For example, temporal profiles 16 of measurement data from the vehicle component or the vehicle can be entered into the control function 28. The control function 28 then calculates the most probable state value 22 of the vehicle component with comparatively few mathematical operations.

[0040] For example, control function 28 can predict the battery voltage of a battery of a battery-electric vehicle based on measured values from the battery or other components of the battery-electric vehicle. An autonomous driving system of the battery-electric vehicle can use the status values to generate corresponding electrical control commands and, for example, schedule charging stops during a trip, issue a warning, and / or book a workshop appointment.

[0041] If the control function 28 is used in the design of a battery electric vehicle, for example, it can be predicted how the battery will behave in certain driving situations in the vehicle.

[0042] Since only a control function 28 is used instead of a neural network, which requires very high computing resources to be executed, only low computing resources are required.

[0043] The example described above does not limit the invention in any way. Rather, the invention can be modified in many ways. All of the features of the invention described above can be essential to the invention alone or in combination with one another. List of reference symbols 10 Database 12 data sets 14 Names of the measurement data 16 temporal courses of the measurement data 18 values of the measured data 20 learning algorithms 22 Condition value 24 internal parameters 26 Software tools 28 Control function

Claims

[1] Computer-implemented method (100) for determining a control function (28) for predicting at least one state value (22) of a vehicle component, the method (100) comprising at least the following steps: a. Determining (102) at least one internal parameter (24) of a learning algorithm (20) which has at least one neural network with a fluid time constant, so that it is designed as an LTC network, and which has been trained by changing the at least one internal parameter (24) to provide at least one output signal indicating at least one state value (22) of the vehicle component from at least one input signal indicating at least one parameter of the vehicle component (14-18); and b. Determining (104) a control function (28) for predicting the state value (22) of the vehicle component by means of a symbolic regression based at least on the input signal and the internal parameter (24) of the trained learning algorithm (20). [2] Computer-implemented method (100) according to claim 1, characterized by that the learning algorithm (20) has at least one neural circuit. [3] Computer-implemented method (100) according to claim 1 or 2, characterized by that the parameter of the vehicle component (14-18) has at least one feature of at least one labeled real measured value from the vehicle. [4] Computer-implemented method (100) according to claim 3, characterized by that at least one characteristic is an element of the group: Input of a control logic, output of a control logic, mathematical operator, name of a variable and / or physical unit. [5] Computer-implemented method (100) according to one of the preceding claims, characterized by that the method (100) further comprises at least the following step before determining the at least one internal parameter (24) of the trained learning algorithm (20): a. Training (106) a learning algorithm (20) with a training data set which has at least measurement data of at least one parameter of the vehicle component as a training input signal and measurement data of the at least one state value (22) of the vehicle component, which are assigned to the measurement data of the parameters, as a target output signal, wherein during training the at least one internal parameter (24) of the learning algorithm (20) is changed in order to adapt an output signal of the learning algorithm (20) based on the training input signal to the target output signal. [6] Computer-implemented method (100) according to one of the preceding claims, characterized by that after training the learning algorithm (20), the internal parameters (24) are fixed (108). [7] Computer-implemented method (100) according to one of the preceding claims, characterized by that the vehicle component is a battery of a battery-electric vehicle and the state value (22) is a battery voltage. [8] Computer-implemented method (100) according to one of the preceding claims, characterized by that the step of determining (104) a control function (28) is carried out using a software tool (26) for performing a symbolic regression. [9] Computer-implemented method (100) according to one of the preceding claims, characterized by that the determined control function (28) is used (110) during operation of the vehicle and / or during construction of the vehicle to predict the at least one state value (22) of the vehicle component. [10] A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method (100) according to any one of claims 1 to 9.

Images